Every run, step by step.
Hand work to an AI agent and you do not have to sit and watch it. It runs in the background. Afterward you can read back exactly what it did: every model turn, every tool it reached for, and how long each one took. When it reaches something it cannot take back, it stops there rather than deciding on your behalf.
A name that outlives its keys
An AI agent keeps one identity however many times its keys are replaced. Its owner is recorded beside it, and a switch is checked every time it asks for anything.
- Turn the switch off and the next request is refused. Nothing is undone behind you, and nothing is left ambiguous.
- Rotating a key does not create a new AI agent, so a year of work stays attached to the same name.
- Every request it makes carries both names: the AI agent, and the person it belongs to.
An AI agent you write starts unapproved
An agent is a model, a set of instructions, a list of what it may use, and a version number. It is listed from the moment it exists and approved by a person, never by itself.
- A new version has to score at least as well as the one already running. That is measured at the pessimistic end of the result, not the flattering one.
- Or a named human overrides that in writing, with the reason, on the record.
- Nothing runs on a project unless it is on that project's approved list.
A stopped run has not failed
Work is queued and done in the background. When a run reaches something it cannot take back, it parks and keeps everything it has done so far. It picks up from the same place once a person decides.
- A refusal is not a dead end either. The AI agent is told, and can take another route.
- The whole run is readable afterward: every model turn, every tool call, how long each took, and where it waited.
- Reversible steps never park. Only what cannot be undone waits for anybody.
Instructions, per project, then shared once approved
An AI agent works from an instruction file somebody wrote. It can be attached to one project and taken off again. It reaches the shared library only through an approval. The reviewer sees a fresh security scan on the request, not one from whenever it was last checked.
Playbooks
A playbook is what a project is measured against: which controls apply, which checks have to pass in each phase, and what evidence closes each one. Apply one to a project and the project is scored against it from its own records as the work happens, so a gap shows while there is still time to close it.
- Change the playbook and every run under it changes with it, from the next run on.
- Runs already on the record stay as they were.
Register your first AI agent with us on a call.
Also in Studio
- StudioAgent securityA name, a limit, and somebody who can stop it.Read the page
- StudioAgent identityA person lets an AI agent in, and can stop it again.Read the page
- StudioWhat an AI agent may doIt never reaches further than its owner.Read the page
- StudioAgent sign-insA sign-in expires sooner the more it can do.Read the page
- StudioApproved modelsWhich models may run, who approved each one, and what it left behind.Read the page
- StudioScans and approved packagesApproved packages at the exact version, watched after that.Read the page
- StudioCommand lineThe terminal your engineers already use, on your rules.Read the page
- StudioMCP serverConnect any AI agent, and the rules come with it.Read the page
- StudioSDK and APIOne typed client for the whole platform, given to your team.Read the page